Publication
Towards Construction Progress Estimation Based on Images Captured on Site
Peter Hevesi; Ramprasad Chinnaswamy Devaraj; Matthias Tschöpe; Oliver Petter; Janis Nikolaus Elfert; Vitor Fortes Rey; Marco Hirsch; Paul Lukowicz
In: EAI IndustrialIoT 2020 - 4th EAI International Conference on Industrial IoT Technologies and Applications. EAI International Conference on Industrial IoT Technologies and Applications (EAI IndustrialIoT-2020), December 11, Online-Conference, Springer, 2020.
Abstract
State of the art internet of things (IoT) and mobile moni- toring systems promise to help gathering real time progress information from construction sites. However, on remote sites the adaptation of those technologies is frequently difficult due to a lack of infrastructure and often harsh and dynamic environments. On the other hand, visual inspection by experts usually allows a quick assessment of a project’s state. In some fields, drones are already commonly used to capture aerial footage for the purpose of state estimation by domain experts.
We propose a two-stage model for progress estimation leveraging im- ages taken at the site. Stage 1 is dedicated to extract possible visual cues, like vehicles and resources. Stage 2 is trained to map the visual cues to specific project states. Compared to an end-to-end learning task, we intend to have an interpretable representation after the first stage (e.g. what objects are present, or later what are their relationships (spa- tial/semantic)). We evaluated possible methods for the pipeline in two use-case scenarios - (1) road and (2) wind turbine construction.
We evaluated methods like YOLOv3-SPP for object detection, and com- pared various methods for image segmentation, like Encoder-Decoder, DeepLab V3, etc. For the progress state estimation a simple decision tree classifier was used in both scenarios. Finally, we tested progress es- timation by a sentence classification network based on provided free-text image descriptions.